I have questions! Does it really require all that hardware? I've spent the last week working on a P25 scanner with 2 RTL-SDRs on a Pi and transcription running on a Ryzen AI laptop. The Pi keeps up with recording fine but it is only 4-5Mhz bandwidth instead of the ~11Mhz you're monitoring. For me transcription only lags by a few seconds but if I ran it on my desktop with a 2080ti it wouldn't even be that much.

What STT model are you using? In my tests I've gotten the best WER with whisper v3 large but it's still not that good.

I would expect that the hardware involved is massive overkill.

The Ettus SDRs are very nice and are marked up enormously over the BOM cost. You’re paying for an integrated, supported system, with good engineering, not for the parts. These days, for this particular use, you could probably build your own direct (no downconversion) receiver with a decent filter and, say, a 400MSPS DAC. (This was not a thing at any reasonable price point 15 years ago when I last played with this stuff.) Or you can buy or build a little mixer and choose an appropriate LO frequency and use a much cheaper DAC. Or you can buy a cheaper SDR.

~15 years ago I amused myself for about an hour using a USRP (older Ettus gear) and MATLAB to do more or less this but for regular FM radio, one channel at a time, on a decent laptop of the era. The processing power needed to channelize and demodulate a few channels from I/Q data is very, very low. Even coming up with some kind of optimal demodulation scheme wouldn’t use many cycles. (Although I suspect that there is more space to try to jointly optimize demodulation and transcription with FM than with AM.)

It would be a fun project to try to kludge up an ESP32 of some sort to do this. The tricky bit would be that the WiFi/BT hardware doesn’t go to DC.

Is P25 not encrypted?